DeepSeek V4 Drops — 1.6 Trillion Parameters, Open Source, 34x Cheaper Than GPT-5.5
Axotron Take
DeepSeek just made frontier AI almost free. Built without Nvidia hardware, open source, and priced at $0.14 per million tokens. The cost argument for closed models is getting harder to make.
Full Notes
DeepSeek V4 is a 1.6 trillion parameter open-source model built without Nvidia hardware, released at $0.14 per million tokens — roughly 34 times cheaper than GPT-5.5 on output pricing.
The "without Nvidia hardware" part is the real story. DeepSeek built a frontier model while under US chip export restrictions, using alternative hardware and architecture optimizations that the Western AI industry largely dismissed as impossible at this scale.
On benchmarks, V4 leads competitive coding with a 3,206 Codeforces Elo rating — the highest of any model, open or closed. It's not the best at everything, but at this price point it doesn't need to be.
For startups and builders, this changes the economics of AI products completely. You can now run near-frontier quality inference for almost nothing. The argument for paying premium closed-model prices is narrowing fast. DeepSeek is not a research project anymore — it's a real competitive threat to every Western AI lab's business model.